Machine Learning Surrogate Model for Accelerating InGaAs/GaAs Quantum Well Laser Design

Abstract

A machine-learning surrogate model based on a multi-layer perceptron neural network is used to predict six laser performance metrics for an indium gallium arsenide quantum well laser given epitaxial layer parameters as inputs. The surrogate model was trained on simulation data from a set of 4999 designs which were generated by varying epitaxial layer parameters and simulating each design with Harold to extract performance metrics. The model predicts all 6 chosen metrics with a coefficient of determination, $R^2$, greater than 0.91 on a held-out test set and demonstrates a $\sim 10^6 \times$ speed up compared to full device simulations. The model is used to search across 1 million potential epitaxial designs in the parameter space and find designs that minimize threshold current and internal loss. Using the surrogate, rather than full device simulation, reduces computation time for a million designs from 231 days to $\sim$2 minutes on a standard laptop. It is also demonstrated that use of the Swish activation function, rather than ReLU, together with engineered input features - i.e. derived quantities calculated from the base model inputs - raise the mean $R^2$ from 0.88 to 0.94 without additional training data, and reduce the mean absolute percentage error from $1.78 \pm 0.08 \%$ to $1.17 \pm 0.09 \%$.